From solving puzzles to masterfully playing a game of chess, current artificial intelligence tools have employed algorithms and generative responses to mimic human intelligence in ways that, at times, exceed our own capabilities. Now, the emerging technology might be more closely matching how we, as humans, learn.
Researchers at Brown set out to highlight the similarities between how people and AI models learn, and what it could reveal about the human brain and today’s computational systems. Jake Russin, a postdoctoral research assistant in computer science at Brown and the first author of a new paper published in the Proceedings of the National Academy of Sciences this August, described the study as “the intersection between psychology, neuroscience and AI.”
“In this study, we are thinking about the neural network models that power today’s AI systems like ChatGPT, and what kinds of computational principles they share with human cognition,” Russin said.
The researchers found that AI learning strategies share several analogous pathways with the human brain, opening up opportunities for the exploration of AI-driven engineering, education and pharmaceuticals.
Russin’s research centers on in-context learning, an emerging trend in the field. During in-context learning, the neural network — a model that processes and identifies patterns in data similar to neurons in the human brain — can learn from just a few input examples and infer results on its own.
This differs from old-school AI systems, which employ in-weight learning, a more common strategy in which AI learns from explicit instructions rather than inferred observations, said Ellie Pavlick, an associate professor of computer science and cognitive and psychological sciences, who also leads the AI Research Institute on Interaction for AI Assistants at Brown.
“Neural networks are just a type of computational model, and the thing that’s very characteristic of them is that they primarily are learning through feedback and association,” Pavlick said.
Pavlick, one of Russin’s advisors for the study, used chess as a “classic example” to describe the model.
Older neural networks are good at learning little-by-little, but cannot generalize rules from just a few examples, Michael Frank, professor of cognitive and psychological sciences, director of the Center for Computational Brain Science and Russin’s second advisor, wrote in an email to The Herald.
“You would write down the rules in code and say: ‘This is how you play chess,’” Pavlick said. In traditional neural networks, “The system would (then) plan over those rules in really explicit terms.”
Newer models employ a different learning strategy. Via in-context learning, the AI “might just take as input a chessboard and then — over looking at many, many games — it’ll kind of infer the rules,” she said. “There’s no explicit instantiation of those rules. They learn the structure from lots of examples rather than having much of it built in.”
“You can go to ChatGPT and say: ‘Strawberries are red, and plums are purple. What is a banana?’ And so it can infer from those few examples that you want the color of the fruit, and it can name the color of the fruit,” Russin explained. “It’s not by learning, but by inferring what the user wants.”
The researchers believe that these two different kinds of learning exhibited by neural networks — incremental learning and rule-based learning — mirror how humans learn.
“We just show that when a neural network is capable of those two different kinds of learning, it naturally reproduces a bunch of phenomena that have been observed in human learning in cognitive science, psychology and neuroscience,” Russin said.
“Sometimes we learn more efficiently by extracting rules — for example, which chess pieces are allowed to move in which positions — but other times we need to just practice repeatedly without knowing exactly why we are getting better — many sports and musical skills,” Frank wrote.
The researchers found that both types of learning share similarities with processes that happen in the human brain, Frank said. He added that incremental learning comes when AI systems recognize the errors they make, which is analogous to how humans use working memory in “incremental reward learning.”
In the research, experimental interactions “reproduced various trade offs while also mimicking some of the same underlying neural signals that have been recorded in the brain,” Frank noted.
The new findings provide multiple opportunities for future uses, he added, explaining that engineers may choose to tailor AI “to optimize for human learning depending on whether they are most likely to learn with one system or the other.”
Get The Herald delivered to your inbox daily.
“Or, one can deliberately design educational programs that try to both enhance one’s ability to learn generalizable rules while still providing (sufficiently) challenging examples that induce errors needed to enhance long term retention,” he added.
“If we have better computational models of how people think and learn, and we understand how people think and learn better, we would be able to use that to do something productive,” Pavlick said. “Whether it’s better education, better medications or better interventions for other things.”